3

Density Matrices and Quantum Operations

In this chapter, we introduce the concepts and properties of general quan-

tum states - mixed states, and also the concept of quantum operations on
an open system. Using these, we can present the general strategy of quan-
tum computing, and possible applications to other problems such as quantum
communication and quantum sensing. While the discussion can apply to infi-
nite dimensional system, we will focus on the finite dimensional cases so that
the Hilbert spaces under considerations are identified with the set of complex
column vectors.

3.1 Mixed States and Density Matrices

Recall that if a quantum state is represented by a unit vector |ψ⟩ ∈ Cn , it
is convenient to represent it as a density matrix |ψ⟩⟨ψ| because |ψ⟩ and |ϕ⟩
will generate the same density matrix if and only if |ψ⟩ = eiα |ϕ⟩ for some
α ∈ R so that they represent the same system. It might happen in some
cases that a quantum system under consideration is in the state |ψi ⟩ with a
probability pi . In other words, we cannot say definitely which state the system
is in. Therefore some random nature comes into the description of the system.
This random nature should not be confused with a probabilistic behavior of
a quantum system. Such a system is said to be in a mixed state, while a
system whose vector is uniquely specified is in a pure state. A pure state is
a special case of a mixed state in which pi = 1 for some i and pj = 0 (j ̸= i).
Mixed states arise naturally in physical systems, for example.

ˆ Suppose we observe a beam of totally unpolarized light and measure

whether photons are polarized vertically or horizontally. The measure-
ment outcome of a particular photon is either horizontal or vertical.
Therefore when the beam passes through a linear polarizer, the inten-
sity is halved. The beam is a uniform mixture of horizontally polarized
photons and vertically polarized photons.

ˆ A particle source emits a particle in a state |ψi ⟩ with a probability

pi (1 ≤ i ≤ N ).

59
60 QUANTUM COMPUTING

ˆ Consider an ensemble of identical closed systems in contact with a

reservoir with temperature T . If we pick up one of the members in
the ensemble, it is in a state |ψi ⟩ with energy Ei with a probability
pi = e−Ei /kB T /Z(T ), where Z(T ) = Tre−H/kB T is the partition func-
tion.
In each of these examples, a particular state |ψi ⟩ ∈ H ≡ Cn appears with
probability pi , in which case the expectation value of the observable a is
⟨ψi |A|ψi ⟩, where we assume |ψi ⟩ is normalized; ⟨ψi |ψi ⟩ = 1. The mean value
of a is then given by
XN
⟨A⟩ = pi ⟨ψi |A|ψi ⟩, (3.1)
i=1
where N is the number of available states. Let us introduce the density
matrix (operator) by
N
X
ρ= pi |ψi ⟩⟨ψi |. (3.2)
i=1

Note that Z 7→ Tr (Z) is a linear function on Mn , and for X = (xrs ) ∈ Mm,n

and Y = (yuv ) ∈ Mn,m , we have
X
Tr (XY ) = xrs ysr = Tr (Y X).
r,s

Thus, Eq. (3.1) is rewritten in a compact form as

⟨A⟩ = Tr(ρA). (3.3)
because the left hand side equals
N
X N
X N
X
pj ⟨ψj |A|ψj ⟩ = pj Tr [⟨ψj |A|ψj ⟩] = pj Tr (|ψj ⟩⟨ψj |A)
i=1 j=1 j=1
N
X
= Tr [( pj (|ψj ⟩⟨ψj |)A] = Tr (ρA).
j=1

Here we use the fact that Tr [µ] = µ ∈ C for the first equality, and Tr (XY ) =
Tr (Y X) for the second equality.
Suppose we have two mixed states obtained by convex combinations of pure
states |ψ1 ⟩⟨ψ1 |, . . . , |ψr ⟩⟨ψr | and |ϕ1 ⟩⟨ϕ1 |, . . . , |ϕs ⟩⟨ϕs | and end up as the same
density matrix ρ, i.e., for two probability vectors (p1 , . . . , pr ), (q1 , . . . , qs ),
r
X s
X
ρ= pj |ψj ⟩⟨ψj | = qj |ϕj ⟩⟨ϕj |.
j j=1

Then the mean value for any observable will be the same. So, we will not
be able to distinguish the two states; practically, they are the same. So, we
Density Matrices and Quantum Operations 61

FIGURE 3.1
Space of density matrices Dn (gray oval) as a subset of unit-trace Hermitian
matrices (white square). The boundary of Dn represents pure states while
the interior represents mixed states that are not pure.

will focus only on the density matrices when we consider a system in mixed
states. We will denote by Dn the set of density matrices in Mn .
Figure 3.1 depicts the space of density matrices Dn (the gray oval) which
is a subset of unit-trace Hermitian matrices (the white square). The state ρ
is represented by a mixture ρ = (1 − t)|ψ⟩⟨ψ| + t|ϕ⟩⟨ϕ|, where t ∈ [0, 1], of
two pure states as well as a mixture of two mixed states ρ = (1 − s)ρ1 + sρ2 .
A general unit-trace Hermitian matrix h may be also represented as h =
(1 + t)ρ3 − tρ4 but the coefficient must be negative in this case.
√
EXAMPLE 3.1.1. Let p1 = p2 = 1/2 and |ψ1 ⟩ = (1, 0)t , |ψ2 ⟩ = (1, 1)t / 2.
The density matrix of this state is
 
1 31
ρ= .
4 11
√

This density matrix also represents a mixed state with q1 = 2 + 2 /4, q2 =
√
√ q √
√ q √
2 − 2 /4 and |ϕ1 ⟩ = (1+ 2, 1)t / 2 2 + 2 , |ϕ2 ⟩ = (1− 2, 1)t / 2(2 − 2).
This mixture is obtained by the spectral decomposition of ρ and hence
⟨ϕ1 |ϕ2 ⟩ = 0.
We may also decompose ρ into two mixed states as ρ = 21 ρ1 + 43 ρ2 , where
   
1 10 1 52
ρ1 = , ρ2 = .
2 01 6 21

Properties which a density matrix ρ satisfies are very much like axioms for
pure states. ∗

∗ The postulates work for trace class operators in infinite dimensional Hilbert space H.
62 QUANTUM COMPUTING

A1′ A physical state of a system, whose Hilbert space is Cn , is completely

specified by its associated density matrix ρ : Cn → Cn . A density
matrix is a positive semi-definite Hermitian operator with Tr ρ = 1 (see
remarks below).
A2′ The mean value of an observable a is given by

⟨A⟩ = Tr (ρA). (3.4)

A3′ The temporal evolution of the density matrix of a closed system is given
by the Liouville-von Neumann equation,
d
iℏ ρ = [H, ρ], (3.5)
dt
where H is the system Hamiltonian (see remarks below).

Several remarks are in order.

ˆ The set {|ψ1 ⟩, . . . , |ψN ⟩⟩} associated with the density matrix (3.2) may
not be orthonormal although ⟨ψj |ψj ⟩ = 1 for each j. Nevertheless, ρ is
Hermitian since pi ≥ 0, and it is positive semi-definite
X X
⟨ϕ|ρ|ϕ⟩ = pi ⟨ϕ|ψi ⟩⟨ψi |ϕ⟩ = pi |⟨ψi |ϕ⟩|2 ≥ 0.
i i

We also have
X X
Tr ρ = ⟨ek |ρ|ek ⟩ = ⟨ek |pi |ψi ⟩⟨ψi |ek ⟩
k i,k
!
X X X
= pi ⟨ψi | |ek ⟩⟨ek | |ψi ⟩ = pi ⟨ψi |ψi ⟩ = 1,
i k i

where {|ek ⟩} is an orthonormal basis of H.

ˆ Each |ψi ⟩ follows the Schrödinger equation
d
iℏ |ψi ⟩ = H|ψi ⟩
dt
in a closed quantum system. Its Hermitian conjugate is
d
−iℏ ⟨ψi | = ⟨ψi |H.
dt
We prove the Liouville-von Neumann equation from these equalities as
d d X X X
iℏ ρ = iℏ pi |ψi ⟩⟨ψi | = pi H|ψi ⟩⟨ψi |− pi |ψi ⟩⟨ψi |H = [H, ρ].
dt dt i i i
Density Matrices and Quantum Operations 63

It is easy to verify that Dn is a convex set, i.e., rρ1 + (1 − r)ρ2 with r ∈ [0, 1]
for ρ1,2 ∈ Dn is also a density matrix. Pn
Note that one can always do a spectral decomposition ρ = j=1 λj |λj ⟩⟨λj |
so that it is a convex combination of its eigenprojections.
EXAMPLE 3.1.2. A pure state |ψ⟩ is a special case in which the corre-
sponding density matrix is
ρ = |ψ⟩⟨ψ|. (3.6)
Therefore ρ in this case is nothing but the projection operator onto the state
|ψ⟩. Observe that
X X
⟨A⟩ = Tr ρA = ⟨ei |ψ⟩⟨ψ|A|ei ⟩ = ⟨ψ|A|ei ⟩⟨ei |ψ⟩ = ⟨ψ|A|ψ⟩.
i i

Let us consider a beam of photons. We take a horizontally polarized state

|0⟩ = | ↔⟩ and a vertically polarized state |1⟩ = | ↕⟩ as orthonormal basis
vectors. If the photons are a totally uniform mixture of two polarized states,
the density matrix is given by
 
1 1 1 10 1
ρ = |0⟩⟨0| + |1⟩⟨1| = = I.
2 2 2 01 2
This state is a uniform mixture of | ↕⟩ and | ↔⟩ and called a maximally or
uniformly mixed state. √
If photons are in a pure state |ψ⟩ = (|0 ⟩ + |1⟩)/ 2, the density matrix,
with {|0⟩, |1⟩} as basis, is
 
1 11
ρ = |ψ⟩⟨ψ| = .
2 11
If |ψ⟩√ itself is used as a basis vector, the other vector being |ϕ⟩ = (|0 ⟩ −
|1⟩)/ 2, the density matrix with respect to the basis {|ψ⟩, |ϕ⟩} has a component
expression  
10
ρ= .
00
Verify that they all satisfy Hermiticity, positive semi-definitness and Tr ρ = 1.
More generally, if we let |x⟩ = (x1 , . . . , xn )t and |y⟩ = (y1 , . . . , yn )t be a
pair of orthonormal vectors in Cn so that |vj ⟩ = (xj , yj )t ̸= 0Pfor every j. Let
n
ρj = p1j |vj ⟩⟨vj | with pj = ∥|vj ⟩∥2 for j = 1, . . . , n. Then ρ = j=1 p2j ρj = 21 I2
is the maximally mixed state.
P
Let A = aPλa |λa ⟩⟨λa | be the spectral decomposition of an observable
A and let ρ = i pi |ψi ⟩⟨ψi | be an arbitrary state. Then the measurement
outcome of A is λa with the probability
X
p(a) = pi |⟨λa |ψi ⟩|2 = ⟨λa |ρ|λa ⟩ = Tr (Pa ρ), (3.7)
i
64 QUANTUM COMPUTING

where Pa = |λa ⟩⟨λa | is the projection operator. The state changes to a pure
state |λa ⟩⟨λa | immediately after the measurement with the outcome λa . This
change is written as ρ 7→ Pa ρPa /p(a).
The following determines when ρ represents a pure state.
THEOREM 3.1.3. A state ρ ∈ Dn is pure if and only if any one of the
following condition holds.
(a) ρ2 = ρ. (b) Tr ρ2 = 1.

Proof. Suppose ρ = |ψ⟩⟨ψ| is a pure state. Then ρ2 = (|ψ⟩⟨ψ|)(|ψ⟩⟨ψ|) =

|ψ⟩⟨ψ| = ρ. Thus, the condition (a) holds. If (a) holds,Pthen Tr ρ2 = Tr ρ = 1.
n
Thus, the condition (b) holds. If (b) holds, and ρ = j=1 λj |λj ⟩⟨λj |, where
Pn 2
Pn
λ1 ≥ · · · ≥ λn ≥ 0 and j=1 λj = 1. Then ρ = λ2 |λ ⟩⟨λ | has
2 2 2
Pn j j j 2
j=1
eigenvalues λ1 , . . . , λn . So, if Tr ρ = 1 = Tr ρ, then 0 = j=1 (λj − λj ) =
Pn
j=1 λj (1 − λj ) so that
Pn
all the nonnegative numbers λj (1 − λj ) is zero. Thus,
λj ∈ {0, 1}. Since j=1 λj = 1, we see that λ1 = 1 and λj = 0 for j > 1.
Thus, ρ = |λ1 ⟩⟨λ1 | is a pure state.

3.2 Uncorrelated, separable and inseparable states

We classify mixed states of a multipartite system into three classes, namely,
uncorrelated, separable and inseparable states; see the definition below. We
use a bipartite system in the definition, but generalization to multiparti-
tle systems should be obvious. Recall that |ψ⟩ = |ψ1 ⟩ ⊗ |ψ2 ⟩ ∈ Cm ⊗ Cn
is a tensor product state in the bipartite system Cm ⊗ Cn . We have
|ψ⟩⟨ψ| = |ψ1 ⟩⟨ψ1 | ⊗ |ψ2 ⟩⟨ψ2 | ∈ P
Mm ⊗ Mn . More generally, if ρ1 ∈ Dm has
m
a spectral decomposition ρ1 = j=1 λj |λj ⟩⟨λj | and ρ2 ∈ Dn has a spectral
Pn
decomposition ρ2 = j=1 µj |µj ⟩⟨µj |, then
X
ρ1 ⊗ ρ2 = λr µs |λr µs ⟩⟨λr µs ⟩.
r,s

Note that Mm ⊗ Mn ≡ MN whenever mn = N . For ρ ∈ Dmn , we will write

ρ ∈ Mm ⊗ Mn to emphasize that ρ is in the bipartite system composed of
subsystems with (mixed) states in Mm and Mn .
DEFINITION 3.2.1. A state ρ ∈ Mm ⊗ Mn is called uncorrelated if it
is written as
ρ = ρ1 ⊗ ρ2 , (3.8)
with ρ1 ∈ Dm , ρ2 ∈ Dn . It is called separable if it is written in the form
X
ρ= pj ρ1,j ⊗ ρ2,j , ρ1,j ∈ Dm , ρ2,j ∈ Dn , (3.9)
j
Density Matrices and Quantum Operations 65
P
where 0 ≤ pi ≤ 1 and i pi = 1. It is called inseparable if ρ does not admit
the decompostion (3.9),

It is important to note that while not every density matrix ρ ∈ Dmn is sep-
arable, i.e., a convex combination of uncorrelated quantum states ρ1,j ⊗ ρ2,j ,
it is always possible to express ρ as a real linear combination of uncorrelated
quantum states with negative coefficients. See Exercise 3.9.
It is also worthwhile to realize that only inseparable states have quantum
correlations analogous to that of an entangled pure state. However, it does
not necessarily imply separable states have no non-classical correlation. It
was pointed out that useful non-classical correlation exists in the subset of
separable states [24].
It is easy to determine whether a given mixed state ρ ∈ Mm ⊗ Mn is an
uncorrelated state. Let ρ = (Prs )1≤r,s≤m , where Prs ∈ Mn . σ2 = Pjj /Tr (Pjj )
for any Pjj ̸= 0. Then ρ is a uncorrelated state if and only if Prs = ars σ2
with ars ∈ C for all 1 ≤ r, s ≤ m. If ars does exist for all 1 ≤ r, s ≤ m, then
ρ = σ1 ⊗ σ2 with σ1 = (ars ) ∈ Mm . See Exercise 3.1

3.3 Partial Trace and Purification

Let A ∈ Mmn ≡ Mm ⊗ Mn . The partial trace of A over Cn is matrix in
Mm defined as
X
A1 = Tr 2 A ≡ (I ⊗ ⟨fk |)A(I ⊗ |fk ⟩) (3.10)
k

where {f1 , . . . , fn } is an orthonormal basis for Cn . It is easy to see that

Tr 2 (A1 ⊗ A2 ) = (Tr A2 )A1 , and Tr 2 is the unique linear Pmap from Mmn to
Mm satisfying A1 ⊗ A2 7→ (Tr A2 )A1 . As a result, if A = P cj A1,j ⊗ A2,j for
some A1,j ∈ Mm , A2,j ∈ Mn and cj ∈ C, then Tr 2 (A) = j cj (Tr A2,j )A1,j .
Consequently, Tr 2 (A) is the same for any choice of orthonormal basis, see
Exercise 3.6.
Suppose ρ is a bipartite state. If we are interested only in the first system
and have no access to the second system, then the partial trace Tr 2 (ρ) allows
us to “forget” about the second system. In other words, the partial trace
quantifies our ignorance of the second system.
The situation is particularly interesting when ρ = |ψ⟩⟨ψ| ∈ Mmn is a
density matrix of a pure state |ψ⟩. To be more concrete, let us consider the
2-qubit state
1
|ψ⟩ = √ (|0⟩|0⟩ + |1⟩|1⟩).
2
66 QUANTUM COMPUTING

The corresponding density matrix is

 
1 0 0 1
1 0 0 0 0
ρ=  .
2 0 0 0 0
1 0 0 1

The partial trace of ρ over the second Hilbert space is

 
X 1 10
ρ1 = Tr 2 ρ = (I ⊗ ⟨i|)ρ(I ⊗ |i⟩) = . (3.11)
2 01
i=0,1

Note that a pure state |ψ⟩ is mapped to a maximally mixed state ρ1 .

Observe that
Tr (ρ1 A) = Tr (ρ(A ⊗ I)) (3.12)

for an observable A acting on the first Hilbert space. The expectation value
of A ⊗ I under that state ρ is equally obtained by using ρ1 .

We have seen above that the partial trace of a pure state density matrix of
a bipartite system over one of the constituent Hilbert spaces yields a mixed
state. How about the converse? Given a mixed state density matrix, is
it always possible to find a pure state density matrix whose partial trace
over the extra Hilbert space yields the given density matrix? The answer
is yes
Pand the process to find a pure state is called the purification. Let
s
ρ = k=1 pk |ψk ⟩⟨ψk | ∈ Dm , where |ψk ⟩⟨ψk | is a pure state for each k. Then
we can let
s
X √
|Ψ⟩ = pk |ψk ⟩ ⊗ |ek ⟩, (3.13)
k=1

where {|e1 ⟩, . . . , |es ⟩} is an orthonormal basis for Cs . We find

√ 
Tr 2 |Ψ⟩⟨Ψ| = (I ⊗ ⟨ei |) pj pk |ψj ⟩|ej ⟩⟨ψk |⟨ek | (I ⊗ |ei ⟩)
X
= pk |ψk ⟩⟨ψk | = ρ. (3.14)
k

Now, for Pany mixed state ρ ∈ Mm , one can always use its spectral decomposi-
r
tion ρ = j=1 λj |λj ⟩⟨λj |, where r is the number of positive eigenvalues of ρ.
Thus it is always possible to purify a mixed state by tensoring an extra Hilbert
space of dimension equal to the rank of ρ. It is easy to see, by construction,
that purification is far from unique. In fact, there are an infinite number of
purifications of a given mixed state density matrix; again see Exercise 3.6.
Density Matrices and Quantum Operations 67

3.4 Quantum Operations for open systems

3.4.1 Operator sum representation and Kraus operators
Recall that a quantum operation on a closed quantum system has the form
Φ : Mn → Mn such that

Φ(A) = U AU † for all A ∈ Mn .

Here, the unitary matrix U may be time-dependent, i.e., U = U (t). In general,

one has to consider open quantum systems, i.e., quantum systems interacting
with other quantum systems; see [2] and [9] for some general background.
This is unavoidable because of the following.

1. The quantum system will always interact with the environment.

2. In quantum computing one often introduce auxiliary system to the pro-

cess to help the computing process.

As a result, if one considers a quantum state ρ ∈ Dn corresponding to the

principal system, i.e., the system one is interested in, there is always another
quantum state, say, σ ∈ Dk corresponding to the environment or the auxiliary
system, such that σ⊗ρ is regarded as the initial state of the total system, which
is a closed system. Then the evolution and transformation of the bipartite
system will be described by unitary similarity transforms

σ ⊗ ρ 7→ U (σ ⊗ ρ)U † .

However, one cannot or need not have complete control of the environment
(auxiliary) system. So, one will apply a partial trace operation to U (σ⊗ρ)U ∗ ∈
Mm ⊗ Mr , where nk = mr, to obtain τ ∈ Mm for our investigation. This
will be the general quantum operations one can apply to a quantum system.
Under this mathematical framework, we have the following theorem.

THEOREM 3.4.1. For every quantumP operation Φ : Mn → Mm there exist

r ∈ N and F1 , . . . , Fr ∈ Mm,n such that j=1 Fj† Fj = In and
r

r
Fj AFj†
X
Φ(A) = for all A ∈ Mn . (3.15)
j=1

Proof. Let Φ : Mn → Mm be a quantum operation. By the previous

discussion, it can be realized as the partial trace of U (σ ⊗ ρ)U † for a suitable
choice of σ ∈ Dk and U ∈ U(kn), where U may be time dependent, governing
the dynamics of the closed system with initial state σ ⊗ ρ. By purification,
we may assume that σ ∈ Dk is a pure state (1, 0, . . . , 0)t (1, 0, . . . , 0) so that
68 QUANTUM COMPUTING

σ ⊗ ρ = ρ ⊕ 0(k−1)n . Suppose ρ 7→ U (σ ⊗ ρ)U † = (Bpq ), where Bjj ∈ Mm

for j = 1, . . . , r with r = nk/m, and we will apply partial trace to obtain
B11 + · · · + Brr as the image of Φ(ρ). We may let the first n columns of U to
from the matrix  
F1
 .. 
F =  .  ∈ Mrm,n (3.16)
Fr
with Fj ∈ Mm,n for each j, then by block multiplication we have

U (σ ⊗ ρ)U † = F ρF †

so that Bjj = Fj ρFj† for j = 1, . . . , r, and

r
Fj ρFj† .
X
Φ(ρ) = B11 + · · · + Brr =
j=1

Thus, the action of Φ on density matrices has the asserted form. Since the
set of density matrices generate all matrices in Mn , we see that Φ has the
asserted form.

One can reverse the above proof to show that every operator Φ of the form
(3.15) corresponds to a quantum operation Φ : Mn → Mm of an open system
as follows. For the given F1 , . . . , Fr in (3.15), we can form the matrix F in
(3.16). The condition j=1 Fj† Fj = In means that the F has orthonormal
Pr

columns. Hence, we can extend F to a unitary U = [F | F̃ ] ∈ Mmr such that

Φ(ρ) is the partial trace of U (σ ⊗ ρ)U † ∈ Mmr for all density matrices ρ.
Thus, Φ is a quantum operation.
Kraus [8] obtained the result in the context of quantum mechanics. The
quantum operation expressed in the form (3.15) is called the operator
sum representation (OSR) of the quantum operation, and the matrices
F1 , . . . , Fr are called the Kraus operators of the quantum operation Φ. In
fact, a quantum operation can also be viewed as a quantum channel, which
describes the change of quantum states ρ as they go through a quantum device.
In such a context, the Kraus operators are also known as error operators,
and we need to find a quantum operation Ψ known as the recovery channel
such that Ψ ◦ Φ(ρ) = ρ for specific choice of ρ in the code space. This will be
the main topic in Chapter 7.

EXAMPLE 3.4.2. Let U1 , . . . , Ur ∈ U(n) and p1 , . . . , pr be positive numbers

summing up to 1. Then Φ : Mn → Mn defined by
r
pj Uj AUj†
X
Φ(A) = for all A ∈ Mn
j=1
Density Matrices and Quantum Operations 69

is a quantum channel known as the random unitary channel or mixed

unitary channel. It is easy to construct
√a unitary V ∈ Mnr ∈ Mn ⊗ Mr with
p1 U1
the first n columns forming the matrix  ...  with orthonormal columns.
 
√
p r Ur
Then    
A0
Φ(A) = Tr 2 V V† for all A ∈ Mn .
0 0
Alternatively, we may let V = U1 ⊕· · ·⊕Ur ∈ Mnr and σ ∈ Dr be the diagonal
matrix with diagonal entries p1 , . . . , pr . Then A 7→ Tr 2 (V (σ ⊗ A)V † ) =
Pr †
j=1 pj Uj AUj .

3.4.2 Quantum channels and Measurements

930978025
Also, quantum measurements can be viewed as quantum Pnoperations on open
systems. As mentioned before a Hermitian matrix A = j=1 λj |λj ⟩⟨λj | is as-
sociated with an observable. If a state ρ ∈ Dn goes through the measurement
process corresponding to A, the state ρ will “collapse” to one of the pure
states |λj ⟩⟨λj | with a probability Tr (Aρ).
More generally, if Psome eigenvalues of A have multiplicities larger than one,
s
we may write A = j=1 λj Pj , where Pj is the projection operator correspond-
ing to the eigenvalue λj for the distinct eigenvalues λ1 , . . . , λs of A. In such a
case, the projective measurement of ρ under the measurement associated
with A is the quantum operation
X
ρ→ Pj ρPj , (3.17)
j

where pj = Tr (Pj ρPj ) = Tr (ρPj ) and the set {P1 , . . . , Pr } satisfies the com-
pleteness relation j Pj Pj† = j Pj = I. Clearly, the projective measure-
P P
ment is a special case of a quantum operation in which the Kraus operators
are Fj = Pj . Upon measurement of ρ, we get the state p1j Pj ρPj ∈ Dn with a
probability pj .
More generally, for any positive semidefinite matrices Q1 , . . . , Qr ∈ Mn
such that Q1 + · · · + Qr = In , there are M1 , . . . , Mr ∈ Mn such that
Mj† Mj = Qj . The measurement operators are then associated with the quan-
tum operation
r
Mj ρMj†
X
ρ 7→
j=1
1 †
so that ρ will change to the quantum state pj Mj ρMj with a probability
pj = = Tr (ρQj ). The set {Q1 , . . . , Qr } = {M1† M1 , . . . , Mr† Mr }
Tr (Mj ρMj† )
is known as the positive operator-valued measure (POVM).
70 QUANTUM COMPUTING

Note that there is flexibility in the choice or construction of M1 , . . . , Mr . In

fact, one may always change Mj to Uj Mj for unitary matrices U1 , . . . , Ur and
we still have (Uj Mj )† (Uj Mj ) = Qj . Of course, the resulting measurement of
ρ will become p1j Uj Mj ρMj† Uj† with pj = Tr (Uj Mj ρMj† Uj† ) = Tr (ρQj ). Note
that the state after the POVM measurement is not determined due to the
ambiguity caused by the choice of U1 , . . . , Ur . In fact, the choice (construc-
tion) of Mj depends on one’s control of the system as mentioned before. For
Mj ρMj† is associated with the
P
example, if the measurement process ρ 7→
effect of the environment, we can have limited control on M1 , . . . , Mr . On the
other hand, if we create an auxiliary system in the laboratory, we may have
more control on constructing M1 , . . . , Mr . Measurement with {M1 , . . . , Mr }
is called the generalized measurement.
EXAMPLE 3.4.3. Suppose Bob will be given a quantum state chosen from
the linearly independent set of unit vectors {|ψ1 ⟩, . . . , |ψm ⟩}, which may not
be orthonormal. He can construct the following POVM {Q1 , . . . , Qm+1 } such
that he will know for sure that |ψj is sent to him if the measurement of the
received state yields Qj if Qj = |ϕj ⟩⟨ϕj |/m, where ⟨ϕj |ϕ
Pjm ⟩ = 1 and ⟨ϕj |ψi ⟩ = 0
for all i ̸= j for j = 1, . . . , m and Qm+1 = I − j=1 Qj . Evidently, a
measurement of |ψj ⟩⟨ψj | will yield Qj or Qm+1 .
In fact, one can associate a POVM with a projective measurement on a
bipartite system with the principal system as a subsystem by the following
proposition.
PROPOSITION 3.4.4. Let {Q1 , . . . , Qr } ⊆ Mn be a POVM. Then there
are projective measurements {P1 , . . . , Pr } in Mr ⊗ Mn and a unitary V ∈
Mr ⊗Mn such that for j = 1, . . . , r, Qj is the leading n×n submatrix of V † Pj V
equivalently, V1† Pj V1 = Qj if V1 is the first n columns of V . Consequently,
we have the following realization of the POVM as a quantum operation Φ on
an open system
r
1/2 1/2
X
Φ(A) = Qj AQj = Tr 1 (Tr V (E11 ⊗ A)V † ) for all A ∈ Mn .
j=1

Proof. Let {|e1 ⟩, . . . , |er ⟩} be the standard basis for Cr andEjj =|ej ⟩⟨ej |
1/2
Q
 1. 
for j = 1, . . . , r. Let Pj = Ejj ⊗ In ∈ Mr ⊗ Mn , and set V1 =   .. . Then

1/2
Qr
† Pr
V1 V1 = j=1 Qj = In . So, V1 has orthonormal columns and we can extend
V1 to a unitary matrix V ∈ Mr ⊗ Mn . It is easy to verify that
 
1/2
Q1
1/2  . 
[Q1 · · · Q1/2
r ]Pj  ..  = Qj
 
1/2
Qr
Density Matrices and Quantum Operations 71

so that Qj is the leading submatrix of V † Pj V , and

r
1/2 1/2
X
Tr 1 (V (E11 ⊗ ρ)V † ) = Qj ρQj .
j=1

3.5 Partial transpose, Entanglement witness

Suppose ρ is not an uncorrelated state. It is hard to determine whether ρ
is separable or inseparable. In fact, this is an NP-hard problem;† see [14].
Nevertheless, we have the following simple test to identify inseparable states.
Define the partial transpose ρpt of ρ ∈ Dmn with respect to the second
Hilbert space Cn as
ρij,kl → ρil,kj , (3.18)
where
ρij,kl = (⟨e1,i | ⊗ ⟨e2,j |) ρ (|e1,k ⟩ ⊗ |e2,l ⟩).
Here {|e1,k ⟩ : 1 ≤ k ≤ m} is the basis for Cm , while {|e2,k ⟩ : 1 ≤ k ≤ n} is the
basis for Cn . In particular, if ρ = ρ1 ⊗ ρ2 , then ρpt = ρ1 ⊗ ρt2 , and if ρ = (Pij )
with Pij ∈ Mn , then ρpt = (Pijt ).
Now, suppose ρ is separable and has the form (3.9). Then the partial
transpose yields X
ρpt = pj ρ1,j ⊗ ρt2,j . (3.19)
i
t
Note here that ρ for any density matrix ρ is again a density matrix since it
is still positive semi-definite Hermitian with unit trace. Therefore the partial
transposed density matrix (3.19) is another density matrix. It was conjec-
tured by Peres [6] and subsequently proved by the Hordecki family [13] that
positivity of the partially transposed density matrix is a necessary and suf-
ficient condition for ρ to be separable in the cases of C2 ⊗ C2 systems and
C2 ⊗ C3 systems. Conversely, if the partial transpose of ρ of these systems is
not a density matrix, then ρ is inseparable. Instead of giving a proof of the
assertion, we look at the following example.
EXAMPLE 3.5.1. Let us consider the Werner state
 1−p 
4 0 0 0
 0 1+p − p 0 
ρ= 4 2
 0 − p 1+p 0  , (3.20)

2 4
0 0 0 1−p 4

† Interested readers can consult the paper [14] for the definition.
72 QUANTUM COMPUTING

where 0 ≤ p ≤ 1. Here the basis vectors are arranged in the order

|0⟩|0⟩, |1⟩|1⟩, |1⟩|0⟩, |1⟩|1⟩.

Partial transpose of ρ yields

 1−p
− p2

4 0 0
 0 1+p
0 0 
ρpt = 
 0
4
1+p
.
0 4 0 
− p2 0 0 1−p
4

Note that we need to consider off-diagonal matrix elements only when we

partically transpose the matrix. We have, for example,

ρ01,10 = (⟨0| ⊗ ⟨1|) ρ (|1⟩ ⊗ |0⟩)

→ (⟨0| ⊗ ⟨0|) ρ (|1⟩ ⊗ |1⟩) = ρpt
00,11 .

For ρpt to be a physically acceptable state, it must have non-negative eigen-

values. The characteristic equation of ρpt is
 3  
pt p+1 1 − 3p
D(λ) = det(ρ − λI) = λ − λ− = 0.
4 4

There are threefold degenerate eigenvalues λ = (1 + p)/4 and a nondegenerate

eigenvalue λ = (1 − 3p)/4. This shows that ρpt is an unphysical state for
1/3 < p ≤ 1. If this is the case, ρ is inseparable.

Suppose a Hermitian matrix H ∈ Mn has eigenvalues λ1 , . . . , λn summing

up to 1. Then H ∈ Dn if and only if all the eigenvalues of H are nonnegative.
Thus, inseparable states ρ can be determined by non-vanishing negativity
defined as
P
j |λi | − 1
N (ρ) ≡ , (3.21)
2
where λj ’s are the eigenvalues of ρpt .
Negativity is one of the so-called entanglement monotones [14], which also
include concurrence, entanglement of formation and entropy of entanglement.‡

EXAMPLE 3.5.2. It was mentioned above that vanishing negativity is equiv-

alent with separability only for C2 ⊗C2 systems and C2 ⊗C3 systems. A counter

‡ Interested readers can see the definitions and background of these concepts in [14] and its
references.
Density Matrices and Quantum Operations 73

example in a C2 ⊗C4 system has been given in Horodecki [15]. Let us consider
b000 0 b0 0
 
0 b 0 0 0 0 b 0 
0 0 b 0 0 0 0 b 
 
000b 0 00 √0 
 
1 
ρ=  1+b 2
1−b 
 (0 ≤ b ≤ 1), (3.22)
0 0 0 0 2 0 0 2 
7b + 1 
b 0 0 0 0 b 0 0 
 
0 b 0 0 0 0 b 0 
√
2
0 0 b 0 1−b
2 0 0 1+b
2

which is known to be inseparable. The partial transposed matrix with respect

to the second system is
b000 0 00 0
 
0 b 0 0 b 00 0 
0 0 b 0 0 b0 0
 

000b 0 0b 0
 
pt 1  √

ρ = 1−b2
. (3.23)
7b + 1  0 b 0 0 1+b


2 00 2


0 0 b 0 0 b0 0 
 
0 0 0 b 0 0b 0 
√
2
0 0 0 0 1−b
2 00 1+b
2

The eigenvalues of ρpt are

b 2b 2b
0, 0, 0, , ,
7b + 1 7b + 1 7b + 1
√
1 + 14b2 + 9b − 98b4 − 70b3 + 23b2 + 12b + 1
,
2 (49b2 + 14b + 1)
√
1 + 14b2 + 9b + 98b4 − 70b3 + 23b2 + 12b + 1
.
2 (49b2 + 14b + 1)
It√can be shown that the seventh eigenvalue
√ takes the maximum value (25 −
2 10)/130 ∼ 0.144 at b = (47 − 10 10)/31 ∼ 0.496 and the minimum value
0 at b = 0, and hence all the eigenvalues are non-negative for 0 ≤ b ≤ 1 in
spite of inseparability of ρ.
Note that the set of separable states form a convex subset in the set of
bipartite states. Using basic matrix theory (functional analysis), for every
inseparable state ρ ∈ Mm ⊗ Mn , there is linear functional f on Hermitian
matrices in Mmn such that f (ρ) > 0 ≥ f (σ1 ⊗ σ2 ) for all σ1 ∈ Dm , σ2 ∈ Dn .
Since every linear functional on Hermitian matrices in Mmn takes the form
f (X) = Tr (F X) for some Hermitian matrix F ∈ Mmn , we can regard F as an
observable on the bipartite system. The linear functional f or the observable
F associated with it is called the entanglement witness of ρ. So, we have
the following.
74 QUANTUM COMPUTING

THEOREM 3.5.3. Let ρ ∈ Mm ⊗ Mn . Then ρ is inseparble if and only if

there is an entanglement witness F such that

Tr (F ρ) > 0 ≥ Tr (F (σ1 ⊗ σ2 )) for all σ1 ∈ Dm , σ2 ∈ Dn .

It should be remarked that finding an entanglement witness of an insepa-

rable state or showing the nonexistence could be a challenging problem.

3.6 Fidelity
It often happens that one has to compare two density matrices and tell how
much they differ from each other. For instance, an experimentalist may con-
duct an experiment and then compare the resulting quantum state with a
certain existing quantum state. A good measure for this purpose is the fi-
delity defined as follows; see [16].
DEFINITION 3.6.1. Let ρ1 , ρ2 ∈ Dn . Then the fidelity is defined by
 q 2
√ √
F (ρ1 , ρ2 ) = Tr ρ1 ρ2 ρ1 . (3.24)

√ √ √
Here, ρ1 is the positive semi-definite square root of ρ1 , and ρ1 ρ2 ρ1 is
positive semi-definite so that we can take its positive semi-definite square root.
We are going to prove some basic properties of fidelity. To do that we
need to introduce some notation and terminology which will be useful in our
subsequent discussion.
For X ∈ Mm,n , let vec(X) ∈ Cmn be the vector obtained by stacking the
columns of X with the first column on the top, and the last column at the
bottom. Clearly, X 7→ vec(X) is an invertible linear map. One can define
the inverse map from Cmn to Mm,n so that |v⟩ 7→ [v]m,n , where the first m
entries of |v⟩ form the first column of [v]m,n , then next m entries of |v⟩ form
the second column of [v]m,n , etc. Moreover, if we define the inner product

⟨X, Y ⟩ = Tr (X † Y )

on Mm,n , the map |v⟩ 7→ [|v⟩]m,n satisfies ⟨v1 |v2 ⟩ = ⟨[|v1 ⟩]m,n , [|v2 ⟩]m,n ⟩.

pOne may define the inner product norm for A ∈ Mm,n by ∥A∥F =
⟨A, A⟩,§ and the following Cauchy-Schwartz inequality holds (see Section
1.8.3:
|⟨A, B⟩| ≤ ∥A∥F ∥B∥F for any A, B ∈ Mm,n .

§ The norm ∥ · ∥F is also known as the Euclidean norm or Frobenius norm.

Density Matrices and Quantum Operations 75
1/2 1/2
THEOREM 3.6.2. Let ρ1 , ρ2 ∈ Dn . If ρ1 ρ2 has singular values s1 ≥
· · · ≥ sn , then
 2
Xn
F (ρ1 , ρ2 ) = F (ρ2 , ρ1 ) =  sj  ,
j=1

and the following conditions hold.

(1) For any unitary U , F (U ρ1 U † , U ρ2 U † ) = F (ρ1 , ρ2 ).
(2) If ρ1 or ρ2 is a pure state, then F (ρ1 , ρ2 ) = Tr (ρ1 ρ2 ).
(3) We have
F (ρ1 , ρ2 ) ∈ [0, 1].
The equality F (ρ1 , ρ2 ) = 1 holds if and only if ρ1 = ρ2 .
The equality F (ρ1 , ρ2 ) = 0 holds if and only if Tr (ρ1 ρ2 ) = 0, equiva-
lently, σ1r σ2s = 0 for any positive numbers r, s.
1/2 1/2
Proof. Let A = ρ1 ρ2 have a singular value decomposition A = XDY †
where X, Y ∈ Mn are unitary and D is the diagonal matrix with diagonal
entries s1 ≥ · · · ≥ sn ≥ 0. Then
 2
 q 2 √ n
√ √ X
F (ρ1 , ρ2 ) = Tr ρ1 ρ2 ρ1 = [Tr AA† ]2 = [Tr (XDX † )]2 =  sj  ,
j=1

and
 2
 q 2 h √ n
√ √ i2 X
F (ρ2 , ρ1 ) = Tr ρ2 ρ1 ρ2 = Tr A† A = Tr (Y DY † ) =  sj  .
j=1
 
† 1/2 † 1/2 1/2 1/2
U † and A have the



It is clear that U ρ1 U U ρ2 U = U ρ1 ρ2
same singular values. Thus (1) holds.
To prove (2), we may assume that ρ1 is a pure state as F (ρ1 , ρ2 ) = F (ρ2 , ρ1 ).
By (1), to compute F (ρ1 , ρ2 ), we may replace (ρ1 , ρ2 ) by (U ρ1 U † , U ρ2 U † ) and
assume that ρ1 is the diagonal matrix with diagonal entries 1, 0, . . . , 0. If ρ2
has (1, 1) entry µ, then
 q 2
1/2 1/2 √ 2
F (ρ1 , ρ2 ) = Tr ρ1 ρ2 ρ1 = µ =µ

= Tr (ρ1 ρ2 ρ1 ) = Tr (ρ21 ρ2 ) = Tr (ρ1 ρ2 ).

1/2 1/2
Now, we turn to (3). Suppose A = ρ1 ρ2 has singular value decomposi-
tion U DV † , where U, V are unitary matrices, and D is the matrix of singular
values. Let Z = V U † . Then
1/2 1/2
F (ρ1 , ρ2 ) = |Tr (AZ)|2 = |Tr (ρ1 ρ2 Z)|2
76 QUANTUM COMPUTING
1/2 1/2 1/2 1/2
≤ ⟨ρ1 , ρ1 ⟩⟨ρ2 Z, ρ2 Z⟩ = (Tr ρ1 )(Tr ρ2 ) = 1.
The inequality follows from the Cauchy-Schwartz inequality. The equality
holds if and only if ρ1 = µρ2 for some µ ∈ C. Comparing the traces on both
sides, one sees that µ = 1. P
n
Now, 0 = F (ρ1 , ρ2 ) = ( j=1 sj )2 if and only if s1 = · · · = sn = 0, i.e.,
1/2 1/2
A = ρ1 ρ2 = 0. It follows that AA† = 0, and hence
1/2 1/2
0 = Tr (AA† ) = Tr (ρ1 ρ2 ρ1 ) = Tr (ρ1 ρ2 ).
1/2 1/2
Since P = ρ1 and Q = ρ2 are positive semi-definite, P Q = 0 is equivalent
to P r Qs = 0 for any r, s > 0. To see this, suppose P = U D1 U † where
U ∈ Mn is unitary and D1 has diagonal entries λ1 ≥ · · · ≥ λn ≥ 0 such that
λk > 0 = λk+1 . Then 0 = U † P QU = U † P U U † QU = D1 (U † QU ) implies
that the Hermitian matrix U † QU has the form 0k ⊕ C. Thus, U † (P r Qs )U =
D1r (0k ⊕ C s ) = 0, and hence P r Qs = 0. Conversely, if P r Qs = 0, we can
apply the above argument to show that P Q = (P r )1/r (Qs )1/s = 0.

One may see the above properties proved in [16]. In [6], the authors sug-
gested using properties (1) – (3) in the above theorem as the basic require-
ments for any generalization of fidelity measures for comparing mixed states.
We remarked that there are other comparisons between the difference of
two quantum states ρ1 , ρ2 . We name a few examples in the following.

ˆ (The Bures distance) DB (ρ1 , ρ2 ) = 2(1 − F (ρ1 , ρ2 )1/2 ).

p

ˆ (The trace distance) ∥ρ − σ∥Tr = Tr |ρ1 − ρ2 |, where |X| = (X † X)1/2

for any square matrix X.

Note that DB (ρ1 , ρ2 ) arises from the quantity

1/2 1/2 1/2 1/2 1/2 1/2
min ∥ρ1 U − ρ2 V ∥2F = min Tr (ρ1 + ρ2 − ρ1 W ρ2 − ρ2 W † ρ1 ),
U,V ∈U(n) W ∈U(n)
√
which equals 2(1 − F (ρ1 , ρ2 )) = DB (ρ1 , ρ2 ). That is why 2 is in the formula.
One can show
p p
1 − F (ρ1 , ρ2 ) ≤ ∥ρ1 − ρ2 ∥Tr ≤ 1 − F (ρ1 , ρ2 ).

3.7 Entropies
Information carried by a physical system is quantified by various entropies,
both in classical and quantum information theories. Let us start with the
Shannon entropy, which was proposed to quantify classical information.
Density Matrices and Quantum Operations 77

3.7.1 Shannon Entropy

Suppose an event x, such as being hit by a car in 24 hours, happens with
probability p(x). The information content I(x) is defined by

I(x) = − log2 p(x). (3.25)

Note that the base of the log function is 2. We will omit 2 in the rest of
this section to simplify the notation. We use ln for loge . For q(x) = 1/p(x),
− log p(x) = log q(x) is approximately the number of digits required to repre-
sent q(x) in binary number. Therefore the information content I(x) is large
if p(x) is small.
Car accident does not happen very often and we are surprised if we hear
our friend was hit by a car. The information we get by this news is large
(I(x) ≫ 1), which is often associated with a big surprise. In contrast, the
probability of watching a dog in 24 hours is close to 1 in usual circumstances
and we do not get new information by this (I(x) ∼ log 1 ∼ 0) and no one will
be surprised by someone watched a dog. Thus, I(x) quantifies the information
we get as well as our surprise at the information.
Let X be a random variable which takes values in X = {x1 , x2 , . . . , xn }
and let p(xk ) be the probability, with which X takes value xk . We only
consider
Pn the case where X is a finite set. The probabilities sum up to one:
k=1 p(x k ) = 1. The Shannon entropy H(p(x)) of this random variable X is
defined as the average of the information contents over X as
n
X
H(p(x)) = − p(xk ) log p(xk ). (3.26)
k=1

H is positive-semidefinite since 0 ≤ p(xk ) ≤ 1 and − log p(xk ) ≥ 0. We

assume x log x is continuous at x = 0 and put

x log x|x=0 = lim x log x = 0.

x→0

H(p(x)) is the average number of bits to store the information.

Let us work out two extreme cases.
(i) Let p(x1 ) = 1 and p(xk ) = 0 (k ̸= 1). Then H(p(x)) = −1 log 1 = 0.
This is the smallest possible value of H since it is positive-semidefinite.
(ii) Suppose there are n possible values of xk and all these values appear with
the same probability p(xk ) = 1/n. Then H(p(x)) = −n n1 log n1 = log n.
Let us show this is the
P maximal possible value of H. We maximize H
under the constraint k p(xk ) = 1. Namely we consider the extremum
of !
Xn n
X
− p(xk ) log p(xk ) − λ p(xk ) − 1 ,
k=1 k=1
78 QUANTUM COMPUTING

where λ is the Lagrange

P multiplier. Under the variation δp(xk ), we
require δH = − k (log p(xk ) + 1 + λ)δp(xk ) = 0, from which
P we find
p(xk ) = 2−1−λ is independent of k. From the constraint k p(xk ) =
1, we obtain λ = log n − 1 and hence p(xk ) = 1/n. We obtain 0 ≤
H(p(x)) ≤ log n for a general {p(xk )}.

We guess from the above two examples that the Shannon entropy is large
if X is more random and small if X is less random.

EXAMPLE 3.7.1. (1) In coin tossing of a uniform coin, probabilities of

getting heads (H) and tails (T) are pH = pT = 1/2. Then X = {H, T } and
the entropy of this random process is
1 1
H(p(x)) = −2 × log = log 2 = 1,
2 2
which means one bit is required to store the information of H or T.
(2) Suppose the probability of getting H is p and T is 1 − p for a non-uniform
coin. Then
H(p(x)) = −p log p − (1 − p) log(1 − p).
Note that H(p(x)) = 0 if p = 0 or p = 1, namely if there is no randomness
in coin tossing, and H(X) takes maximum value log 2 = 1 at p = 1/2, i.e.,
the most random case.
(3) Probability of getting any of 1, 2, . . . , 6 is 1/6 on a roll of the uniform dice.
The entropy of this process is
1 1
H(p(x)) = −6 log = log 6 ∼ 2.58.
6 6
It requires 3 bits to store the number on average.

Entropy is additive. Let X, Y be two independent random variables taking

values in X = {xj } and Y = {yk } and let p(xj ) and q(yk ) be their probability
distributions. Suppose X and Y are measured and values xj and yk are ob-
tained. This event takes place with probability p(xj )q(yk ). Then the entropy
of this process is
X
H(p(x), q(y)) = − p(xj )q(yk ) log(p(xj )q(yk ))
j,k
X
=− p(xj )q(yk )[log p(xj ) + log q(yk )]
j,k
X X
=− p(xj ) log p(xj ) − q(yk ) log q(yk )
j k
= H(p(x)) + H(q(y)), (3.27)

showing entropy is additive for independent random variables.

Density Matrices and Quantum Operations 79

3.7.2 Classical Relative Entropy

Let X be a classical random variables and X = {xk } be the set of its values.
Suppose there are two probability distributions {p(xk )} and {q(xk )} on the
same variable set X . The classical relative entropy of {p(xk )} to {q(xk )}
is defined by
X
H(p(x)||q(x)) = p(xk )(log p(xk ) − log q(xk )). (3.28)
k

PROPOSITION 3.7.2. H(p(x)||q(x)) ≥ 0, where equality holds if and only

if p(xk ) = q(xk ) for all k.

Proof. Use the well known inequality ln 2 log x = ln x ≤ x − 1 for x > 0 to

show
 
X p(xk ) 1 X q(xk )
H(p(x)||q(x)) = p(xk ) log ≥ p(xk ) 1 −
q(xk ) ln 2 p(xk )
k k
1 X
= [p(xk ) − q(xk )] = 0.
ln 2
k

Clearly the inequality above is saturated if and only if q(xk )/p(xk ) = 1 for all
k, that is p(xk ) = q(xk ) for all k.
Relative entropy measures how close two distributions are. In this sense,
relative entropy is also known as the Kullback-Leibler distance although it
does not satisfy the axioms of distance.

3.7.3 von Neumann Entropy

Quantum counterpart of the Shannon entropy is the von Neumann entropy.
Let us consider a system with the n-dimensional Hilbert space H. Suppose
the system is in a mixed state ρ. Then the von Neumann entropy is defined
as
S(ρ) = −Tr (ρ log ρ), (3.29)
where log stands
P for log2 as before.
LetPρ = i λi |λ i ⟩⟨λi | be the spectral decomposition
Pof ρ, where 0 ≤ λi ≤ 1
and i λi = 1. Now S(ρ) is represented as S(ρ) = − i λi log λi , from which
we find S(ρ) ≥ 0. Since a unitary transformation does not change the eigen-
values of ρ, von Neumann entropy is invariant under unitary transformation;
S(U ρU † ) = S(ρ), U ∈ U(n). (3.30)
Let us consider two extreme cases as before.
(i) Consider a pure state ρ = |ψ⟩⟨ψ|. If we take |ψ⟩ as the first basis
vector of H, its component expression is |ψ⟩ = (1, 0, . . . , 0)t and ρ =
diag(1, 0, . . . , 0), from which we obtain S(ρ) = 0.
80 QUANTUM COMPUTING

(ii) Let us consider next the uniformly mixed state ρ = In /n, where In is
the identity matrix in Mn . Then
1 1
S(ρ) = −n
log = log n,
n n
which is the maximal possible value of S. To prove this, let us extremize
S̃(ρ) = −Tr (ρ log ρ)−µ(Tr ρ−1), where µ is the Lagrange multiplier that
takes care of the constraint Tr ρ = 1. We impose δ S̃(ρ) = −Tr [(log ρ +
1+µ)δρ] = 0 under the variation δρ, from which we obtain ρ = 2−1−µ In .
The unit trace condition gives µ = log n − 1 as before, from which we
obtain ρ = In /n and S(ρ) = log n.
The above observation shows that 0 ≤ S(ρ) ≤ log n for a general ρ. S(ρ)
measures how much ρ differs from pure states. Note that S(ρ) = 0 if and
only if ρ is a pure state. In fact, S(ρ) has been shown to have only one
maximum at ρ = In /n above, namely at λi = 1/n, ∀i, where {λi } is the set
of eigenvalues of ρ. Then the minima are found at the edges, where λk = 1
and λj = 0 (j ̸= k), for 1 ≤ k ≤ n. These edge points are pure states.

3.7.4 Entanglement Entropy

Consider a bipartite system with the associated Hilbert space HAB = HA ⊗HB
with dimHA = nA and dimHB = nB . Take a pure state |ψ⟩ ∈ HAB and ρ =
|ψ⟩⟨ψ| and define the reduced density matrices ρA = Tr B ρ and ρB = Tr A ρ.
PROPOSITION 3.7.3. Suppose there is a pure state |ψ⟩ ∈ HAB = HA ⊗
HB and let ρA and ρB be the reduced density matrices. Then
(i) S(ρA ) = S(ρB ).
(ii) S(ρA ) = S(ρB ) = 0 if and only if |ψ⟩ is a tensor product state.
Pr √ (A) (B)
Proof. (i) Let |ψ⟩ = i=1 si |fi ⟩|fi ⟩ be the Schmidt decomposition of
(a) (a)
|ψ⟩, where r is the Schmidt number of |ψ⟩. Extend {|f1 ⟩, . . . , |fr ⟩} to an
(a) (a) (a) (a)
orthonormal basis {|f1 ⟩, . . . , |fr ⟩, |er+1 ⟩, . . . |ena ⟩} for Ha , a = A and B.
Partial traces of |ψ⟩⟨ψ| with respect these bases yield
r r
(A) (A) (B) (B)
X X
ρA = Tr B |ψ⟩⟨ψ| = si |fi ⟩⟨fi | and ρB = Tr A |ψ⟩⟨ψ| = si |fi ⟩⟨fi |,
i=1 i=1
Pr
from which we obtain S(ρA ) = S(ρB ) = − i=1 si log si .
(ii) If |ψ⟩ = |ψA ⟩|ψB ⟩ be a tensor product state, then |ψ⟩⟨ψ| = |ψA ⟩⟨ψA | ⊗
|ψB ⟩⟨ψB |. It follows ρA = |ψA ⟩⟨ψA | and ρB = |ψB ⟩⟨ψB |, and S(ρA ) =
S(ρB ) = 0.
Conversely, suppose S(ρA ) = S(ρB ) = 0. Then both ρA and ρB are pure
states as we have observed previously. Let us write ρA = |ψA ⟩⟨ψA | and ρB =
|ψB ⟩⟨ψB |. Since we assume the bipartite state is pure, it must take a form
|ψ⟩ = |ψA ⟩|ψB ⟩.
Density Matrices and Quantum Operations 81

Entanglement entropy vanishes when a pure bipartite state is a tensor prod-

uct state. Finiteness of entanglement entropy signals entanglement of the bi-
partite system. In this sense, entanglement entropy works as a measure of
purity or entanglement.

EXAMPLE 3.7.4. In C2 ⊗ C3 , take a vector

1
|ψ⟩ = (i|0⟩|0⟩ + |1⟩|2⟩ + i|1⟩|0⟩ + |2⟩|2⟩).
2
Let us find the entanglement entropies of this state. It is easy to find the
Schmidt decomposition of |ψ⟩ as

1 (1) (2) (1) (2)

|ψ⟩ = √ (|f1 ⟩|f1 ⟩ + |f2 ⟩|f2 ⟩)
2
(a)
We do not need the explicit form of |fi ⟩ here. We find
 
  100
1 10 1
ρ1 = Tr 2 |ψ⟩⟨ψ| = , ρ2 = Tr 1 |ψ⟩⟨ψ| = 0 1 0,
2 01 2
000

(a)
where we have taken {|fi ⟩} as bases. We obtain S(ρ1 ) = S(ρ2 ) = 2 × 1/2 ×
log2 2 = 1.

In contrast with ordinary thermodynamic entropy, entanglement entropy

does not depend on the system size. S(ρA ) = S(ρB ) holds even if A is much
larger than B.
It should be noted that the equality S(ρA ) = S(ρB ) is valid only for
pure bipartite states and does not hold for general mixed states. Suppose
ρAB = ρA ⊗ ρB for example, where ρA and ρB are arbitrary mixed states
of subsystems A and B, respectively. Then S(ρA ) = −Tr ρA log ρA and
S(ρB ) = −Tr ρB log ρB are different in general.

PROPOSITION 3.7.5. The von Neumann entropy and its entanglement

entropy satisfy
(i) S(ρAB ) ≤ S(ρA ) + S(ρB ) (Subadditivity)
(ii) |S(ρA ) − S(ρB )| ≤ S(ρAB ) (Araki-Lieb inequality)
(iii) S(ρAB ) + S(ρBC ) ≥ S(ρABC ) + S(ρB ) (Strong additivity).

The proof of the above inequalities is found in [NC], for example.

3.7.5 Quantum Relative Entropy

DEFINITION 3.7.6. Let ρ, σ ∈ S(H). The relative entropy of ρ to σ is
defined by
S(ρ||σ) = Tr (ρ log ρ) − Tr (ρ log σ). (3.31)
82 QUANTUM COMPUTING

Note that the relative entropy is well defined only when ker ρ ⊇ ker σ so
that log 0 is multiplied by 0. Otherwise S(ρ||σ) is divergent.

PROPOSITION 3.7.7. Let ρ, σ ∈ S(H). Then

(i)
S(ρ||σ) ≥ 0 (Klein’s inequality) (3.32)
The equality holds if and only if ρ = σ.
(ii)
1
S(ρ||σ) ≥ ∥ρ − σ∥21 , (Pinsker’s inequality) (3.33)
2
√
where ∥O∥1 = Tr O† O.
(iii) Suppose ρAB and σAB are states of a bipartite system AB and ρA and
σA are reduced density matrices of ρAB and σAB , respectively. Then

S(ρA ||σA ) ≤ S(ρAB ||σAB ) (Monotonicity). (3.34)

The proof is found in [NC], for example. ∥ρ − σ∥1 is a measure of distin-

guishability of two states ρ and σ. It is hard to distinguish them if ∥ρ − σ∥1 is
small. The relative entropy gives an upper bound of their distinguishability.
As a result, it is harder to distinguish two states if the part B of the bipartite
system is ignored by partial trace, which justifies the monotonicity inequality.
Now we are ready to proceed to the world of quantum information and
quantum computation. Variations on the themes introduced here and in the
previous chapter will appear repeatedly in the following chapters.

3.8 Notes and Open problems

Theorem 3.4.1 followed from a result of Choi [7] as a consequence of completely
positive maps in the context of C ∗ -algebra. There has been new approach to
general quantum operations that do not assume that the initial state of the
principal system and the environment system is a product state to begin with.
For example, see [15].
Using the materials discussed in the first three chapters, we can describe
the general ideas on quantum computing and quantum information research.
For example, quantum information theory concerns the use of quantum
state to store and transmit data securely and efficiently. For example, using
the no-cloning theorem and the fact that measuring quantum states will alter
them, we can design secure quantum encryption system as shown in the fol-
lowing chapter. Since quantum states always interact with the environment
leading to degradation changes of the data (known as decoherence), we need
to design of quantum error correction schemes to fight against such effects. If
Density Matrices and Quantum Operations 83

there is an ensemble of identical quantum state, one may apply measurement

to these identical states, and get an estimate of the density matrix repre-
senting them. This leads to the tomography problem. Similarly, if one can
determine the input and output states of a quantum channel many times, one
may get an estimate of the error operators in the operator sum representation
of the quantum channel. This is the subject of quantum tomography. In par-
ticular, the quantum channel tomography is needed for the design quantum
error correction schemes.
One may give a general description of quantum algorithms for computing
in terms of vector states as follows.
1. Set up the problem using a quantum sate |ψ⟩, say, with n qubits.
2. Construct a bipartite system by setting up |ϕ⟩ ⊗ |ψ⟩.
3. Apply a suitable quantum operation to get |Ψ⟩ = U (|ϕ⟩ ⊗ |ψ⟩).
4. Apply a suitable measurement to |Ψ⟩ to obtain useful information.
Researchers have also applied quantum effects to other branches of sci-
ences such as photosynthesis, evolution, image processing, quantum machine
learning, quantum neural network, quantum cognition, etc. For example, see
[16, 17, 18] and their references. In general, one would use a quantum state to
encode the data needed to be processed in the problem. Then apply a suitable
quantum operation to manipulate quantum state leading to the solution of
the problem. For example, for image recognition problem, suppose one wants
to decide whether a given image corresponds to the number 0, 1, . . . , 9. One
can encode 0, 1 . . . , 9 as orthonormal vector states |ψ0 ⟩, . . . , |ψ9 ⟩. Then design
a quantum operation that will identify correctly a given state |ψ⟩ (correspond-
ing to an imperfect image) to a state in the set {|ψj : 0 ≤ j ≤ 9}, which |ψ⟩
meant to be.
Here are some open problems.
1. Let ρ1 ∈ Dm , ρ2 ∈ Dn . Determine the set
S(ρ1 , ρ2 ) = {ρ ∈ Dmn : Tr 1 (ρ) = ρ1 , Tr 2 (ρ) = ρ1 }.

2. Determine ρ ∈ S(A, B) with maximum rank and minimum rank.

3. Determine ρ ∈ S(A, B) with maximum von Neumann entropy S(ρ) =
Tr (−ρ ln ρ).
4. More generally, one may consider tripartite system with states in
Dn1 n2 n3 and determine S(ρ23 , ρ13 ) = {ρ ∈ Dn1 n2 n3 : Tr 1 (ρ) =
ρ23 , Tr 2 (ρ) = ρ13 }, where ρ23 ∈ Dn2 n3 and ρ12 ∈ Dn1 n3 are two given
states.
5. Given quantum states ρ1 , . . . , ρk ∈ Dn , σ1 , . . . , σk ∈ Dm , determine
quantum operation Φ such that Φ(ρj ) = σj for j = 1, . . . , k.
84 QUANTUM COMPUTING

Exercises for Chapter 3

EXERCISE 3.1. Let ρ = (Pij ) ∈ Dmn with Pij ∈ Mn . Suppose σ2 =
Pjj /Tr (Pjj ) ∈ Mn for some Pjj ̸= 0. Then ρ is an uncorrelated state if and
only if ρrs = ars σ2 with ars ∈ C for all 1 ≤ r, s ≤ m. If ars do exist for all
1 ≤ r, s ≤ m, then ρ = σ1 ⊗ σ2 with σ1 = (ars ) ∈ Mm .
Note: You need to argue why σ1 , σ2 are density matrices.
EXERCISE 3.2. Verify that
 1+p p 
4 0 0 2
 0 1−p 0 0 
ρ1 =  4  (0 ≤ p ≤ 1) (3.35)
 0 0 1−p
4 0 
p 1+p
2 0 0 4

is a density matrix. Show that the negativity does not vanish for p > 1/3.
EXERCISE 3.3. Verify that
p p 
2 0 0 2
 0 1−p 1−p
0
ρ2 =  2 2  (0 ≤ p ≤ 1) (3.36)
 0 1−p 1−p
0
2 2
p p
2 0 0 2

is a density matrix. Show that the negativity vanishes only for p = 1/2.
EXERCISE 3.4. Let
1
|ψ⟩ = √ (|0⟩|1⟩ − |1⟩|0⟩).
2
Find the corresponding density matrix. Then partial-trace it over the first
Hilbert space to find the reduced density matrix of the second system.
EXERCISE 3.5. Let  
1 10
ρ1 =
4 03
be a density matrix with a basis {|0⟩, |1⟩}. Find a purification of ρ1 .
EXERCISE 3.6. Let
s
X √
|Ψ⟩ = pk |ψk ⟩ ⊗ |ϕk ⟩
k=1
Ps
be a purification of ρ1 = k=1 pk |ψk ⟩⟨ψk | ∈ Dm , where {|ϕ1 ⟩, . . . , |ϕs ⟩} is an
orthonormal basis for Cs . Show that
X√
|Ψ′ ⟩ = pk |ψk ⟩ ⊗ U |ϕk ⟩
k

is another purification of ρ1 for any U ∈ U(n).

Density Matrices and Quantum Operations 85

EXERCISE 3.7. Let U be a unitary operator acting on ρ1 and ρ2 . Show

that
F (U ρ1 U † , U ρ2 U † ) = F (ρ1 , ρ2 ). (3.37)

EXERCISE 3.8. Let

   
1 0 0 0 1 0 0 1
1 0 0 0 0 , ρ2 = 1  0
 0 0 0
ρ1 = 
 .
2 0 0 0 0 2 0 0 0 0
0 0 0 1 1 0 0 1

Find the fidelity F (ρ1 , ρ2 ).

EXERCISE 3.9. Show that every Hermitian matrix H ∈ Mmn can be writ-
ten as a linear combination of density matrices of the form ρ1 ⊗ ρ2 with
ρ1 ∈ Dm and ρ2 ∈ Dn using the following steps.
(1) Let {|e1 ⟩, . . . , |em ⟩} be the standard basis for Cm . Consider the set B1
consisting of matrices in Dm of the form: Dj = |ej ⟩⟨ej | for 1 ≤ j ≤ m, or

Xrs = |er + es ⟩⟨er + es |/2, Yrs = |er − ies ⟩⟨er + ies |/2 for 1 ≤ r < s ≤ m.

If H P= (hrs ) is Hermitian and hrs = xrs + iyrs with xrs , yrs ∈ R, show that
H − r<s (xrs Xrs + yrs Yrs ) is a real diagonal matrix, and hence is a linear
combination of D1 , . . . , Dm . Conclude that H is a real linear combination of
the elements in B1 .
Similarly, if {|f1 ⟩, . . . , |fn ⟩} is the standard basis for Cn , then every Hermitian
matrix G = (grs ) ∈ Mn is a real linear combination of the elements in the set
B2 consisting of elements in Dn of the form |fj ⟩⟨fj | for 1 ≤ j ≤ n, or

|fr + fs ⟩⟨fr + fs |/2, |fr − ifs ⟩⟨fr + ifs |/2 for 1 ≤ r < s ≤ n.

(2) Let B = {ρ1 ⊗ρ2 : ρ1 ∈ B1 , ρ2 ∈ B2 }. For any Hermitian T = (Trs )1≤r,s≤m

with Trs ∈ Mn , we have Tjj ∈ Mn is Hermitian, and Trs = Hrs + iGrs with
†
P Hrs = (Trs + Trs )/2, Grs = i(Tij − Tij ) ∈ Mn for r < s.
Hermitian matrices
Show that T1 = r<s (Xrs ⊗ Hrs + Yrs ⊗ Grs ) is a real linear combination of
matrices in B, and deduce that T − T1 is a real linear combination of matrices
in B, and so is T .

EXERCISE 3.10. Let ρ ∈ Dmn .

(a) Suppose ρ is a pure state. Then ρ is separable if and only if ρ = ρ1 ⊗ ρ2
for pure states ρ1 ∈ Dm , ρ2 ∈ Dn .
(b) Suppose ρ is a separable. Then ρ is a convex combination of product
states ρ1 ⊗ ρ2 such that ρ1 ∈ Dm , ρ2 ∈ Dn are pure states.

EXERCISE 3.11. Suppose Φ : Mn → Mm is a quantum operation with the

operator sum representation (3.15)
86 QUANTUM COMPUTING

(a) Show that Φ is trace preserving, i.e., Tr Φ(A) = Tr A for all A ∈ Mn .

[It suffices to check Φ(Eij ) = δij , where {E11 , E12 , . . . , Enn } is the standard
basis for Mn .
(b) Show that Φ maps positive semi-definite matrices to positive semi-
definite matrices.
(c) Show that Φ is k-positive for every positive integer k, i.e., if A = (Aij ) ∈
Mk (Mn ), where Aij ∈ Mn for all 1 ≤ i, j ≤ k, is positive semi-definite, then
so is [Φ(Aij )].
[Note that [Φ(Aij )] = j=1 (Fj ⊗ Ik )A(Fj† ⊗ Ik ).]
Pr

EXERCISE 3.12. Prove the following uncertainty principle for a mixed

state ρ ∈ Mn . Let A,p B ∈ Mn be Hermitian matrices,
p (α, β) =
(Tr (Aρ), Tr (Bρ)), ∆(A) = Tr [(A − αI)2 ρ] and ∆(B) = Tr [(B − βI)2 ρ].
Then
∆(A)∆(B) ≥ |Tr ([A, B]ρ)|/2.
The equality holds if and only if there is θ ∈ [0, 2π) such that

cos θ(A − αI)ρr + i sin θ(B − βI)ρr = 0 (3.38)

for any positive number r.

References
[1] E. Riefell and W. Polak, Quantum Computing: A Gentle Introduction,
MIT Press (2011).
[2] M. A. Neilsen and I. L. Chuang, Quantum Computation and Quantum
Information, Cambridge University Press (2000).
[3] D. M. Greenberger, M. A. Horne and A. Zeilinger, in ’Bell’s Theorem,
Quantum Theory, and Conceptions of the Universe, ed. M. Kafatos,
Kluwer, Dordrecht (1989). Also avilable as arXiv:0712.0921 [quant-ph].
[4] W. Dür, G. Vidal and J. I. Cirac, Phys. Rev. A 62, 062314 (2000).
[5] A. Einstein, B. Podolsky, N. Rosen, Phys. Rev. 41, 777 (1935).
[6] Y.-C. Liang, Y.-H. Yeh, P.E.M.F. Mendonça, R.Y. Teh, M.D. Reid, and
P.D. Drummond, Quantum fidelity measures for mixed states, Reports
on Progress in Physics, Volume 82, Number 7, (2019).
[7] M.-D. Choi, Completely Positive Linear Maps on Complex Matrices,
Linear Algebra and its Applications, 10, 285–290 (1975).
[8] K. Kraus, States, Effects and Operations: Fundamental Notions of
Quantum Theory, Springer Verlag, (1983)
Density Matrices and Quantum Operations 87

[9] K. Hornberger, e-print quant-ph/0612118 (2006).

[10] H. Barnum, M. A. Nielsen and B. Schumacher, Phys. Rev. A 57, 4153
(1998).
[11] Y. Kondo, et al., J. Phys. Soc. Jpn. 76 (2007) 074002.
[12] G. Lindblad, Commun. Math. Phys. 48, 119 (1976).
[13] V. Gorini, A. Kossakowski and E. C. G. Sudarshan, J. Math. Phys., 17,
821 (1976).
[14] Gurvits, L., Classical deterministic complexity of Edmonds’ problem and
quantum entanglement, in Proceedings of the 35th ACM Symposium on
Theory of Computing, ACM Press, New York, 2003.
[15] H. Hayashi, G. Kimura, and Y. Ota,. Kraus representation in the pres-
ence of initial correlations. Physical Review A - Atomic, Molecular, and
Optical Physics, 67(6), 2003. 621091-621095, 2003.
[16] M. Mohseni, Y. Omar, G. Engel, M.B. Plenio, Quantum effects in biol-
ogy, Cambridge University Press, Cambridge, 2014.
[17] M. Asano, A. Khrennikov, M. Ohya, Y. Tanaka, I. Yamato, Quantum
Adaptivity in Biology: From Genetics to Cognition, Springer, New York,
2020.
[18] F. Yan and S.E. Venegas-Andraca, Quantum Image Processing,
Springer, New York, 2010.